Cancer Biology and Survival Analysis in Cancer Trials: Restricted Mean Survival Time Analysis versus Hazard Ratios

Hazard ratios are commonly used when comparing survival between two groups and make the assumption that the relative event rates do not change markedly during follow-up, i.e. that event rates are proportional. However, there is currently debate about the use of the proportional hazards assumption to summarise the treatment effect in survival analysis compared with restricted mean survival time (RMST) analysis, particularly in cancer trials. In many situations it is unrealistic to assume that relative event rates in two groups will be proportional throughout follow-up and, hence, RMST analysis, which does not make this assumption, may be preferable. Several benefits of the latter approach have been identified but the biological perspective is not often discussed. Biological features such as the patterns of tumour growth and response can also contribute to assessing the relative merit of these two methods; such biological considerations are the subject of this paper. The observation that the most commonly observed approximation to the underlying distribution of time to event data, the lognormal distribution, does not reliably show proportional hazards in the comparison of two groups, lends weight to a statistical approach that is not based on proportional hazards. The proportional hazards assumption should be viewed more critically when estimating treatment effects. An optimum approach may be to include both proportional hazards analysis and RMST analysis when comparing time to event endpoints.